Orthogonal subspace projection-based approaches to classification of MR image sequences

Orthogonal subspace projection (OSP) approach has shown success in hyperspectral image classification. Recently, the feasibility of applying OSP to multispectral image classification was also demonstrated via SPOT (Satellite Pour 1’Observation de la Terra) and Landsat (Land Satellite) images. Since an MR (magnetic resonance) image sequence is also acquired by multiple spectral channels (bands), this paper presents a new application of OSP in MR image classification. The idea is to model an MR image pixel in the sequence as a linear mixture of substances (such as white matter, gray matter, cerebral spinal fluid) of interest from which each of these substances can be classified by a specific subspace projection operator followed by a desired matched filter. The experimental results show that OSP provides a promising alternative to existing MR image classification techniques.     

Dialectical multispectral classification of diffusion-weighted magnetic resonance images as an alternative to apparent diffusion coefficients maps to perform anatomical analysis

Multispectral image analysis is a relatively promising field of research with applications in several areas, such as medical imaging and satellite monitoring. A considerable number of current methods of analysis are based on parametric statistics. Alternatively, some methods in computational intelligence are inspired by biology and other sciences. Here we claim that philosophy can be also considered as a source of inspiration. This work proposes the objective dialectical method (ODM): a method for classification based on the philosophy of praxis. ODM is instrumental in assembling evolvable mathematical tools to analyze multispectral images. In the case study described in this paper, multispectral images are composed of diffusion-weighted (DW) magnetic resonance (MR) images. The results are compared to ground-truth images produced by polynomial networks using a morphological similarity index. The classification results are used to improve the usual analysis of the apparent diffusion coefficient map. Such results proved that gray and white matter can be distinguished in DW-MR multispectral analysis and, consequently, DW-MR images can also be used to furnish anatomical information.     

Multi-level adaptive segmentation of multi-parameter MR brain images.

MR brain image segmentation into several tissue classes is of significant interest to visualize and quantify individual anatomical structures. Traditionally, the segmentation is performed manually in a clinical environment that is operator dependent and may be difficult to reproduce. Though several algorithms have been investigated in the literature for computerized automatic segmentation of MR brain images, they are usually targeted to classify image into a limited number of classes such as white matter, gray matter, cerebrospinal fluid and specific lesions. We present a novel model-based method for the automatic segmentation and classification of multi-parameter MR brain images into a larger number of tissue classes of interest. Our model employs 15 brain tissue classes instead of the commonly used set of four classes, which were of clinical interest to neuroradiologists for following-up with patients suffering from cerebrovascular deficiency (CVD) and/or stroke. The model approximates the spatial distribution of tissue classes by a Gauss Markov random field and uses the maximum likelihood method to estimate the class probabilities and transitional probabilities for each pixel of the image. Multi-parameter MR brain images with T(1), T(2), proton density, Gd+T(1), and perfusion imaging were used in segmentation and classification. In the development of the segmentation model, true class-membership of measured parameters was determined from manual segmentation of a set of normal and pathologic brain images by a team of neuroradiologists. The manual segmentation was performed using a human-computer interface specifically designed for pixel-by-pixel segmentation of brain images. The registration of corresponding images from different brains was accomplished using an elastic transformation. The presented segmentation method uses the multi-parameter model in adaptive segmentation of brain images on a pixel-by-pixel basis. The method was evaluated on a set of multi-parameter MR brain images of a twelve-year old patient 48h after suffering a stroke. The results of classification as compared to the manual segmentation of the same data show the efficacy and accuracy of the presented methods as well as its capability to create and learn new tissue classes.     

Evaluation of Alzheimer's disease by analysis of MR images using multilayer perceptrons and committee machines.

Alzheimer's disease is the most common cause of dementia, yet hard to diagnose precisely without invasive techniques, particularly at the onset of the disease. This work approaches image analysis and classification of synthetic multispectral images composed by diffusion-weighted magnetic resonance (MR) cerebral images for the evaluation of cerebrospinal fluid area and its correlation with the advance of Alzheimer's disease. The MR images were acquired from a unique volunteer with Alzheimer's, using an image system based on a clinical 1.5T tomographer. The classification methods are based on multilayer perceptrons and committee machines and the classification results are used to correlate clinical and imaging findings. The classification results are used to improve the usual analysis of the ADC map.     

Segmentation of MR brain images into cerebrospinal fluid spaces, white and gray matter

An interactive computer method for quantifying CSF, white matter, and gray matter in magnetic resonance (MR) axial brain scans is presented. A stripping algorithm is used to remove the skull and scalp from each axial section. The images are then filtered to correct for radiofrequency inhomogeneity image artifacts. Late echo images are subtracted from or added to early echo images to enhance fluid/tissue and gray/white tissue contrast, respectively. Thresholds for fluid/tissue and gray/white separation are set interactively. A boundary pixel locking algorithm is used to handle ambiguities due to partial voluming between the fluid and tissue compartments. The MR brain scans from five healthy, young, normal men were obtained using a standard neuroanatomical reference technique. These data were processed and percentages computed for fluid, gray matter and white matter compartments. The gray/white ratios compare favorably with those determined in a published postmortem brain study.     

Multispectral analysis of MR images of the breast

Preliminary investigations were conducted into the potential of magnetic resonance (MR) images for tissue classification of the breast on the basis of relative signal intensity. Multispectral techniques originally developed by the National Aeronautics and Space Administration for satellite image analysis were used in sequence selection, image data correction, image standardization, and image interpretation. Numerous sequence combinations with varying repetition times (TR) and echo times (TE) were considered, and a triplet was selected consisting of long TR/long TE, short TR/short TE, and an opposed phase sequence with intermediate TR and TE. Correction to remove system-imposed intensity inhomogeneities was required for all images. Image standardization based on fat and pectoral muscle signals was necessary for intercase comparisons. Multispectral images obtained based on this analysis suggest the feasibility of intensity-based image classification.     

Multispectral magnetic resonance image analysis using principal component and linear discriminant analysis

PURPOSE: To explore the possibilities of combining multispectral magnetic resonance (MR) images of different patients within one data matrix. MATERIALS AND METHODS: Principal component and linear discriminant analysis was applied to multispectral MR images of 12 patients with different brain tumors. Each multispectral image consisted of T1-weighted, T2-weighted, proton-density-weighted, and gadolinium-enhanced T1-weighted MR images, and a calculated relative regional cerebral blood volume map. RESULTS: Similar multispectral image regions were clustered, while dissimilar multispectral image regions were scattered in a single plot. Both principal component and linear discriminant analysis allowed discrimination between healthy and tumor regions on the image. In addition, linear discriminant analysis allowed discrimination between oligodendrogliomas and astrocytomas. However, the discriminant analysis method was partially capable of recognizing the tumor identity in unknown multispectral images. CONCLUSION: The proposed method may help the radiologist in comparing multispectral MR images of different patients in a more easy and objective way.     

Improving image contrast using principal component analysis for subsequent image segmentation

This article presents a technique for improving MR image contrast by linearly combining multiple MR images with different tissue contrast. The weighting coefficients of the linear combination are derived using principal component analysis. The contrast-enhanced composite image is segmented subsequently using gray level-based 1D segmentation methods. The technique reduces a multispectral image set to composite eigenimages and allows application of appropriate 1D segmentation methods that do not have equivalent counterparts in multispectral methods.     

Automated classification of multi-spectral MR images using Linear Discriminant Analysis

Magnetic resonance imaging (MRI) is a valuable instrument in medical science owing to its capabilities in soft tissue characterization and 3D visualization. A potential application of MRI in clinical practice is brain parenchyma classification. This work proposes a novel approach called “Unsupervised Linear Discriminant Analysis (ULDA)” to classify and segment the three major tissues, i.e. gray matter (GM), white matter (WM) and cerebral spinal fluid (CSF), from a multi-spectral MR image of the human brain. The ULDA comprises two processes, namely Target Generation Process (TGP) and Linear Discriminant Analysis (LDA) classification. TGP is a fuzzy-set process that generates a set of potential targets from unknown information, and applies these targets to train the optimal division boundary by LDA, such that three tissues GM, WM and CSF are separated. Finally, two sets of images, namely computer-generated phantom images and real MR images are used in the experiments to evaluate the effectiveness of ULDA. Experiment results reveal that UDLA segments a multi-spectral MR image much more effectively than either FMRIB's Automated Segmentation Tool (FAST) or Fuzzy C-means (FC).     

Mathematical segmentation of grey matter, white matter and cerebral spinal fluid from MR image pairs

The aims of this study were (1) to design a mathematical segmentation technique to allow extraction of grey matter, white matter and cerebral spinal fluid volumes from paired high resolution MR images and (2) to document the statistical accuracy of the method with different image combinations. A series of linear equations were derived that describe proportional tissue volumes in individual image voxels. The equations use estimates of pure tissue values to derive the proportion of each tissue within a single voxel. Repeatability of manual estimations of pure tissue values was assessed both using regions of interest and thresholding techniques. Statistical accuracy of tissue estimations for a variety of image pairs was assessed from measurements of root-mean-square noise and mean grey level intensity. The technique was used to produce parametric images of grey and white matter distribution. The segmentation technique showed greatest statistical accuracy when the first image has high grey/white matter contrast and the second image has little contrast or the rank order of the signal intensities from pure tissue is reversed. A combination of inversion recovery fast spin echo and fast FLAIR images produced a statistical error of 11% for grey matter and 10% for white matter for any given voxel. The effect of increasing sample size improves both of these figures to give a 1% statistical error on a 100 pixel sample.     

Local histogram correction of MRI spatially dependent image pixel intensity nonuniformity

We describe a computationally straightforward post-hoc statistical method of correcting spatially dependent image pixel intensity nonuniformity based on differences in local tissue intensity distributions. Pixel intensity domains for the various tissues of the composite image are identified and compared to the distributions of local samples. The nonuniformity correction is calculated as the difference of the local sample median from the composite sample median for the tissue class most represented by the sample. The median was chosen to reduce the effect ers on determining the sample statistic and to allow a sample size small enough to accurately estimate the spatial variance of the image intensity nonuniformity. The method was designed for application to two-dimensional images. Simulations were used to estimate optimal conditions of local histogram kernel size and to test the accuracy of the method under known spatially dependent nonuniformities. The method was also applied to correct a phantom image and cerebral MRIs from 15 healthy subjects. Results show that the method accurately models simulated spatially dependent image intensity differences. Further analysis of clinical MR data showed that the variance of pixel intensities within the cerebral MRI slices and the variance of slice volumes within individuals were significantly reduced after nonuniformity correction. Improved brain-cerebrospinal fluid segmentation was also obtained. The method significantly reduced the variance of slice volumes within individuals, whether it was applied to the native images or images edited to remove nonbrain tissues. This statistical method was well behaved under the assumptions and the images tested. The general utility of the method was not determined, but conditions for testing the method under a variety of imaging sequences is discussed. We believe that this algorithm can serve as a method for improving MR image segmentation for clinical and research applications.     

Pharmacokinetic parameters in CNS Gd-DTPA enhanced MR imaging

Dynamic MR imaging can be used to study tissue perfusion and vascular permeability. In the present article a procedure for dynamic MR is presented, which (a) accurately resolves the fast kinetics of tissue response during and after intravenous infusion of the paramagnetic contrast medium Gd-DTPA and (b) yields a linear relationship between the measured MR signal and the Gd-DTPA concentration in the tissue. According to these features, the measured signal-time curves can be analyzed within the framework of pharmacokinetic modeling. Tissue response has been parameterized using a linear two-compartment open model, with only negligible effects of the peripheral compartment on the central compartment. The three model parameters were fitted to the signal-time data pixel by pixel, based on a set of 64 rapid SE images (SE 100/10 ms, image scan time 13 s, interscan intervals 11 s). This makes it possible to construct parameter images, whereby structures become visible that cannot be distinguished in conventional Gd-DTPA enhanced MR. As a clinical example, the approach is discussed in a case of glioblastoma.     

Frequency encoding for simultaneous display of multimodality images.

An original method for simultaneous display of functional and anatomic images, based on frequency encoding (FE), merges color PET with T1-weighted MR brain images, and grayscale PET with multispectral color MR images. A comparison with two other methods reported in the literature for image fusion (averaging and intensity modulation techniques) was performed. METHODS: For FE, the Fourier transform of the merged image was obtained summing the low frequencies of the PET image and the high frequencies of the MR image. For image averaging, the merged image was obtained as a weighted average of the intensities of the two images to be merged. For intensity modulation, the red, green and blue components of the color image were multiplied on a pixel-by-pixel basis by the grayscale image. A comparison of the performances of the three techniques was made by three independent observers assessing the conspicuity of specific MRI and PET information in the merged images. For evaluation purposes, images from seven patients and a computer-simulated MRI/PET phantom were used. Data were compared with a chi-square test applied to ranks. RESULTS: For the depiction of MRI and PET information when merging color PET and T1-weighted MR images, FE was rated superior to intensity modulation and averaging techniques in a significant number of comparisons. For merging grayscale PET with multispectral color MR images, FE and intensity modulation were rated superior to image averaging in terms of both MRI and PET information. CONCLUSION: The data suggest that improved simultaneous evaluation of MRI and PET information can be achieved with a method based on FE.     

Digital staining for multispectral images of pathological tissue specimens based on combined classification of spectral transmittance

In this study, the digital transformation (digital staining) of the 16-band multispectral image of a hematoxylin and eosin (HE) stained pathological specimen to its Masson's trichrome (MT) stained counterpart is addressed. The digital staining procedure involves the classification of the various H&E-stained tissue components and then the transformation of their transmittance spectra to their equivalent MT-stained transmittance configurations. Combination of transmittance classifiers were designed to classify the various tissue components found in the multispectral images of an HE-stained specimen, e.g. nucleus, cytoplasm, red blood cell (RBC), fibrosis, etc.; while pseudo-inverse method was used to obtain the transformation matrices that would translate the transmittance spectra of the classified HE-stained multispectral pixels to their MT-stained configurations. To generate the digitally stained image, weighting factors, which were based on the classifiers beliefs, were introduced to the generated transformation matrices. Initial results of our experiments on liver specimens show the viability of multispectral imaging (MSI) to implement a digital staining framework in the pathological context.     

Segmentation strategies for breast tumors from dynamic MR images

This paper describes two semiautomated methods of segmentation of breast tumors from dynamic MR images obtained subsequent to administration of gadopentate dimeglumine. The first method, based on temporal correlation, generates a similarity map from the dynamic scans in which the value of each pixel is determined by its temporal similarity to a reference region of interest. The second method uses multispectral analysis and generates a feature map from a scatterplot of pixel intensities in the pre- and postcontrast images. The segmentation methods were tested on malignant and benign breast lesions in 11 patients with a range of tumor volumes and percentage contrast enhancement. The accuracy of both segmentation techniques and reproducibility of the multispectral method were investigated. A comparison of the two methods established that the temporal correlation method was superior based on accuracy, extent of user interaction, and speed of segmentation.     

MR image contrast at high field strength.

The T1 of soft tissues increases with magnetic field strength. Some tissue contrast may be diminished on high-field-strength magnetic resonance (MR) images when conventional TRs are used, because of altered T1 effects on the MR signals. This necessitates longer TRs in techniques that use long TRs, which prolongs the examination excessively. Behavior of macroscopic magnetization is governed by the Bloch equations. Therefore, T1 contributions to the MR signal can be modulated by means of both timing intervals and radio-frequency pulses. The analytic solution to the Block equations allowed calculation of white matter/gray matter and gray matter/cerebrospinal fluid contrast in both spin-echo and inversion-recovery (IR) imaging. Rabbit brains (normal and tumor-containing) were then imaged in vivo at 1.5 and 4.7 T. In addition, MR images of a human head were obtained at 4.0 T. Experimental results supported the theoretical predictions that brain contrast on long TR spin-echo or IR images increases with field strength. However, varying the excitation flip angle allowed optimization of the T1 contribution to the MR signals, improving image contrast and/or reducing examination time. Thus, the dependence of T1 on field strength determines the optimum choice of imaging techniques and parameters in a predictable fashion.     

Cine MR angiography of the heart with segmented true fast imaging with steady-state precession

In five healthy subjects and 18 patients, cine magnetic resonance (MR) imaging of the heart was performed with a true fast imaging with steady-state precession (FISP) sequence. Results were compared both quantitatively and qualitatively with those at cine fast low-angle shot (FLASH) MR imaging. The blood-myocardial contrast-to-noise ratio (CNR) was 2.0 times higher and the normalized (for measurement time and pixel size) blood-myocardial CNR was 4.0 times higher for true FISP compared with FLASH MR imaging. Qualitative scores for image quality were significantly higher with true FISP MR imaging. Segmented cine true FISP MR imaging generated high-contrast MR images of the heart in healthy subjects and in patients with heart disease and produced image quality superior to that with cine FLASH MR imaging.     

Cerebral hemodynamics and cerebral blood volume: MR assessment using gadolinium contrast agents and T1-weighted Turbo-FLASH imaging.

PURPOSE: To assess the degree and regional pattern of first-pass brain enhancement using dynamic MR imaging. MATERIALS AND METHODS: Ultrafast MR imaging (1.06-second acquisition time per image) was performed in 19 healthy subjects following a bolus IV injection of a gadolinium contrast agent; 36 patients with suspected pathology were studied using the same protocol. RESULTS: Calculated percent blood volumes were 4.9% for right cortical gray matter, 4.8% for left cortical gray matter, and 2.6% for white matter. Subtraction images were obtained that depicted the first pass "blood pool" pattern of enhancement (gray and white matter) which was significant. CONCLUSION: Preliminary evidence suggests utility for cerebral "blood pool" imaging, especially if reduced image acquisition times can be achieved.     

Diffusion weighted MR: principles and clinical use in selected brain diseases

PURPOSE: To define the principles and technical bases of diffusion weighted MR imaging of the brain and report our experience in the evaluation of selected brain disorders including age-related ischemic white matter changes (leukoaraiosis), neoplastic and infective cysts and wallerian degeneration. MATERIAL AND METHODS: Between May 1999 and June 2000 we examined seventeen patients: 10 patients with leukoaraiosis and deterioration of cognitive and motor function, 5 patients with focal cystic lesions (one anaplastic astrocytoma, one glioblastoma, one metastasis from squamous cell lung carcinoma, one pyogenic abscess and one case with cerebral tubercolosis) and 2 patients with wallerian degeneration (one with post-hemorrhagic degeneration of right corticospinal tract and one with post-traumatic degeneration of left optic tract). All patients underwent a standard cranial MR examination including SE T1-, proton density, T2-weighted, FLAIR and diffusion weighted images. Post-contrast T1-weighted sequences were also obtained in the patients with cystic lesions. Diffusion weighted images were acquired with double shot echoplanar sequences. Diffusion sensitizing gradient along the x, y and z axes and b values ranging 800 to 1200 s/mm2 were used. For each slice a set of three orthogonal diffusion "anisotropic" images, an "isotropic" image and a standard T2-weighted image were reconstructed. Postprocessing included generation of the apparent diffusion coefficient maps and of the "trace" image that reflects pixel by pixel the diffusional properties of water particles only. Values of mean diffusivity within regions of interest were computed in the "trace" image and compared with those obtained in contralateral brain areas. In patients with leukoaraiosis the diffusivity in posterior periventricular white matter was compared with that measured in 10 age-matched control subjects without leukoaraiosis. RESULTS: In patients with leukoaraiosis the areas of increased periventricular signal intensity on T2-weighted images showed a significantly higher (p < 0.001) diffusivity (mean values 124.7 +/- 21.3 x 10(-5) mm2/s) as compared to control subjects (mean values 85 +/- 7 x 10(-5) mm2/s). Diffusion weighted images in 2 patients revealed the presence of a small focal area of increased signal and reduced diffusivity in "trace" images consistent with recent ischemic lesion. In neoplastic cystic lesions the central necrotic/cystic content was always hypointense on diffusion weighted images and showed increased diffusivity on "trace" images. On the other hand the central necrotic content of the pyogenic brain abscess was hyperintense and showed low diffusivity. In patients with wallerian degeneration diffusion weighted images and "trace" images demonstrated loss of anisotropy and increased diffusivity in the affected white matter tract relative to the contralateral. DISCUSSION: The increased diffusivity observed in areas of leukoaraiosis and the identification of subclinical acute ischemic lesions by diffusion weighted images might be more useful than standard MR sequences for monitoring the disease progression. Diffusion weighted images allow differentiation of the different parts of focal cystic lesions (edema, solid and cystic/necrotic portion) and are useful to differentiate pyogenic brain abscess from necrotic tumors. In patients with wallerian degeneration the loss of anisotropy and the increase of diffusivity values in the affected tract are probably related to myelin breakdown and allow better recognition of the affected tract relative to standard MR images. CONCLUSIONS: Diffusion weighted MR imaging can be performed during a standard cranial MR examination and add useful clinical information in several brain disorders besides acute ischemic stroke.     

In vivo MR tractography using diffusion imaging

Diffusion in structured tissue, such as white matter or muscle, is anisotropic. MR diffusion tensor imaging (DTI) measures anisotropy per pixel and provides the directional information relevant for MR tractography or fiber tracking in vivo. MR tractography is non-invasive, relatively fast, and can be repeated multiple times without destructing important tissue. Moreover, the combination with other MR images is relatively simple. In this paper, the basic principles of tractography are presented. Different tracking methods with varying degrees of complexity are introduced and their potential strengths and weaknesses are discussed. Clinical applications and different strategies for evaluating the fidelity of tracking results are reviewed.